Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

Generalized Inverse Kinematics via Conditional Flow

Kaitlyn Zhang, Tomás Lozano-Pérez, Leslie Pack Kaelbling

November 2026

GenGenIK solutions for a simulated RB-Y1 at a door and reaching across a cluttered table.

TL;DR GenGenIK trains one conditional flow-matching Transformer to sample whole-body IK solutions for mobile bases, multiple end-effectors, multi-target paths, observed scenes and posture preferences. On five simulated RB-Y1 path tasks it solves 92.7-100% of problems with 32 samples in 20-36 ms, against 22.0-92.0% and 167-1067 ms for cuRobo.

What they do

Planners need many IK solutions, not one: alternative base placements, starts from which a whole path stays reachable, collision-free postures in clutter. GenGenIK samples whole-body configurations, or trajectories over several targets, conditioned on end-effector targets, a scene point cloud and optional preferences. All experiments are in simulation.

How it works

  • Masked flow matching. An encoder-decoder Transformer maps noise to base and joint values. Masks set which targets are active and which coordinates are fixed, so one architecture covers point, multi-end-effector, mobile-base and path IK. A soft bounding function keeps outputs inside joint limits.
  • Data. cuRobo solutions certified with Drake, from 27 MolmoSpaces kitchens and about 2,500 procedural environments. Scenes enter as a 4096-point labeled cloud encoded into 32 tokens.
  • Speed. Fine-tuning through the sampler with collision and stability penalties, reflow distillation to one step, then a light Levenberg-Marquardt refiner.
  • Preferences. Non-differentiable requests (base stand-off, elbow height) are conditioning tokens; differentiable ones (manipulability, stability) use guidance kept in the constraint tangent space.

Results

  • Grasp IK on RB-Y1 in 252 held-out scenes: 80.9% success@1 and 98.2% success@8, versus 29.0% and 58.1% for cuRobo. From a partial cloud: 61.4% and 90.1%, versus 12.0% and 52.1%. With no cloud or another scene’s cloud, success@1 falls to 16.7% and 20.7%.
  • Path IK (door/drawer, pushing, bimanual lift, flip-up, under-table pickup): 92.7-100% success@32 in 20-36 ms to a first valid solution, versus 22.0-92.0% and 167-1067 ms for cuRobo. Conditioning on the full path makes 58.2% of first configurations extendable to the whole task, versus 24.3%.
  • Scene-free point IK: lower position and rotation error than IKDiffuser on FR3, a LEAP hand and G1, with fewer network evaluations. Guidance on G1 raises manipulability 28.42% (IKDiffuser 24.69%) and cuts stability violations 68.74%.

Limitations

No real-robot experiments. A model trained without under-table pickup scores 0% on it, so it does not extrapolate to new kinds of scene interaction, and each task family still needs expert-solver data. In scene-free point IK, cuRobo has lower latency under large GPU batches.

Figures

From Figure 1: Example use cases of GenGenIK: under-table pickup, partially observable scenes and cluttered kitchens.
From Figure 1: Example use cases of GenGenIK: under-table pickup, partially observable scenes and cluttered kitchens.
From Figure 2: Conditional architecture. Task poses and scene geometry form conditioning tokens, while the current ca...
From Figure 2: Conditional architecture. Task poses and scene geometry form conditioning tokens, while the current candidate is represented by base and joint tokens. The decoder combines these with forward-kinematics feedback and optional solution requests.
From Figure 3: Preference-conditioned IK in the RB-Y1 bimanual lift task. Left: requesting a base stand-off of at mos...
From Figure 3: Preference-conditioned IK in the RB-Y1 bimanual lift task. Left: requesting a base stand-off of at most 0.45 m from the anchor target shifts sampled base placements inward compared with unconditional samples. Right: requesting the right elbow at least 0.15 m above its shoulder and the left elbow at least 0.10 m below gives an asymmetric posture.
Abstract

Inverse kinematics (IK) is a foundational primitive in robotics, mapping the task-space objectives expressed by higher-level planners to robot configurations that realize them. Traditional IK methods generally treat each query as an isolated pose-to-configuration problem, but planners often need a distribution over valid realizations that accounts for future reachability and preferences such as collision avoidance and stability. We introduce GenGenIK, a generative modeling approach that supports mobile bases, multiple end-effectors and targets, scene awareness, and preferences over robot behavior within a common interface. Experiments show that GenGenIK outperforms existing methods in both solve rate and runtime across tasks such as bimanual lift and flip-up, while allowing flexible control over the sampled distribution and remaining effective under partial scene observations. Website: https://icra-2027-gengenik.github.io/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.